On Mon, 1 Jun 2020 at 10:00, Al Grant Al.Grant@arm.com wrote:
(Hi - posted for review. There was a recent ABI breaking change (see commit reference below), after which perf inject on ETM creates corrupt files. The simple fix ends up with it using the new ABI. I have an alternative patch set that emulates the old ABI - for the benefit of tools like autofdo that can't yet consume the new ABI - but it's much messier.)
From: Al Grant al.grant@arm.com
Commit 42bbabed09ce6208026648a71a45b4394c74585a ("perf tools: Add hw_idx in struct branch_stack") changed the format of branch stacks in perf samples. When samples use this new format, a flag must be set in the corresponding event. Synthesized branch stacks generated from CoreSight ETM trace were using the new format, but not setting the event attribute, leading to consumers seeing corrupt data. This patch fixes the issue by setting the event attribute to indicate use of the new
format.
Fixes: 42bbabed09ce ("perf tools: Add hw_idx in struct branch_stack") Signed-off-by: Al Grant al.grant@arm.com Acked-by: Andrea Brunato andrea.brunato@arm.com
diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 62d2f9b9ce1b..71a056e29675 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -1332,7 +1332,13 @@ static int cs_etm__synth_events(struct
cs_etm_auxtrace *etm,
} if (etm->synth_opts.last_branch)
{ attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;/* We don't use the hardware index, but the sample generationcode uses the new format branch_stack with this field,so the event attributes must indicate that it's present. */attr.branch_sample_type |=- PERF_SAMPLE_BRANCH_HW_INDEX;
I'm fine with this set and would ask you to send it to the public mailing list when 5.8-rc1 has been released. Out of interest I have the following 2 questions:
Would setting sample.no_hw_idx to 'true' would have the same effect?
Unfortunately not. The purpose of no_hw_idx seems to be to modify the layout of the branch_stack structure in the expanded perf_sample structure, and in the ETM code, that does have the hw_idx field. If we set no_hw_idx hen we have to change the layout of sample->branch_stack, i.e. remove hw_idx, or use perf_sample__branch_entries throughout.
And then synthetic-events.c produces new-format data regardless of no_hw_idx:
if (type & PERF_SAMPLE_BRANCH_STACK) { sz = sample->branch_stack->nr * sizeof(struct branch_entry); /* nr, hw_idx */ sz += 2 * sizeof(u64); memcpy(array, sample->branch_stack, sz); array = (void *)array + sz; }
We would have to change it to something like
if (type & PERF_SAMPLE_BRANCH_STACK) { *array++ = sample->branch_stack->nr; if (!sample->no_hw_idx) *array++ = sample->branch_stack->hw_idx; sz = sample->branch_stack->nr * sizeof(struct branch_entry); memcpy(array, perf_sample__branch_entries((struct perf_sample *)sample), sz); array = (void *)array + sz; }
It's possible, but the patch ends up being ten times bigger (I tried) and still doesn't fix the technical debt. perf_sample__branch_entries is just wrong. There is no need for non-standard layouts in the unpacked perf_sample or branch_stack. What matters is (a) parsing serialized data according to flags in the event attributes and (b) writing serialized (synthetic) data that matches the flags in the event attributes. Both of those suggest having access to the event attributes when we're writing data, instead of passing in a random subset of values like 'type'.
Have you tried with PT? I see that Kan's patch applied the same fix to both cs- etm.c and intel-pt.c and as such I would expect to find the same problem there as well.
Yes, now you mention it, I see the same problem with PT, fix is same:
diff --git a/tools/perf/util/intel-pt.c b/tools/perf/util/intel-pt.c index 23c8289c2472..334e59238f65 100644 --- a/tools/perf/util/intel-pt.c +++ b/tools/perf/util/intel-pt.c @@ -2896,8 +2896,10 @@ static int intel_pt_synth_events(struct intel_pt *pt,
if (pt->synth_opts.callchain) attr.sample_type |= PERF_SAMPLE_CALLCHAIN; - if (pt->synth_opts.last_branch) + if (pt->synth_opts.last_branch) { attr.sample_type |= PERF_SAMPLE_BRANCH_STACK; + attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX; + }
if (pt->synth_opts.instructions) { attr.config = PERF_COUNT_HW_INSTRUCTIONS;
Al
Thanks, Mathieu
} if (etm->synth_opts.instructions) { attr.config = PERF_COUNT_HW_INSTRUCTIONS; IMPORTANTNOTICE: The contents of this email and any attachments are confidential and
may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
CoreSight mailing list CoreSight@lists.linaro.org https://lists.linaro.org/mailman/listinfo/coresight
IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.